3, 297–305
This paper presents an analysis of some class of bilinear systems that can be applied to biomedical modelling. It com-
bines models that have been studied separately so far, taking into account both the phenomenon of gene amplification and
multidrug chemotherapy in their different aspects. The mathematical description is given by an infinite dimensional state
equation with a system matrix whose form allows decomposing the model into two interacting subsystems. While the first
one, of a finite dimension, can have any form, the other is infinite dimensional and tridiagonal. A methodology of the anal-
ysis of such models, based on system decomposition, is presented. An optimal control problem is defined in the l1 space.
In order to derive necessary conditions for optimal control, the model description is transformed into an integro-differential
form. Finally, biomedical implications of the obtained results are discussed.
most of the existing forms of therapy consist in using 2. A cell of Type i for i ≥ 1 may mutate in a short
several drugs instead of a single one. Then, modelling time interval (t, t + dt) into a Type i + 1 cell with
should take into account increasing drug resistance to each probability bi dt + o(dt) and into a Type i − 1 cell
chemotherapeutic agent used. with probability di dt + o(dt). A cell of Type i = 0
Finally, the phase-specific control of the drug- may mutate in a short time interval (t, t + dt) into
sensitive cancer population will be addressed. Actually, a Type 1 cell with probability α dt + o(dt), where
each drug affects cells being in particular phases and it α is several orders of magnitude smaller than any of
makes sense to combine these drugs so that their cu- bi ’s and di ’s.
mulative effect on the cancer population is the greatest. 3. The drug action results in a fraction ui of ineffective
So far, phase-specific chemotherapy has been considered divisions in the cells of Type i (hence 0 ≤ ui ≤ 1).
without any regard to problems stemming from increas-
ing drug resistance. Combining the infinite dimensional 4. The process is initiated at time t = 0 by a finite
model of drug resistance with the phase-specific model population of cells of different types.
of chemotherapy should bring mathematical modelling
much closer to its clinical application. A graph representing the possible flows between sub-
populations is presented in Fig. 1(a).
So far, only the simplest case has been investigated, realistic, it is sensible to find the optimal control which
in which the resistant cells are completely insensitive to minimises the performance index
the drug’s action and there are no differences between the
l−1 ∞ m Z T
parameters of cells of different types: X X X
J = Ni (T )+r1 Ni (T ) + r uk (τ ) dτ , (4)
i=0 i=l k=0 0
Ṅ0 (t) = 1 − 2u(t) λN0 (t) − αN0 (t) + dN1 (t),
where r1 and r are non-negative weighting factors.
Ṅ1 (t) = λN1 (t) − (b + d)N1 (t) + dN2 (t)
The idea on which such optimisation is based is to
+αN0 (t),
(2) minimise the resistant cancer subpopulation at the end of
.
.. the therapy while minimizing the negative cumulative ef-
fect of the drug represented by the integral component.
Ṅi (t) = λNi (t) − (b + d)Ni (t) + dNi+1 (t)
+bNi−1 (t) for i ≥ 2.
3. Model Taking into Account the Partial
However, using the same line of reasoning that has Sensitivity of the Resistant Subpopulation
been applied to that case, it is also possible to analyse a In this case, it is assumed that the resistant subpopulation
less simplified model. If it is assumed that the parameters consists of two parts: the one, which is partially sensi-
may vary for a given finite number of cells and are the tive to the drug, and the other, which is completely drug-
same only for the infinite dimensional tail of the system, resistant. Then the following set of equations is obtained:
the following model can be investigated:
Ṅ0 (t) = 1−2u(t) λ0 N0 (t)−αN0 (t)+d1 N1 (t),
Ṅ0 (t) = 1 − 2u0 (t) λ0 N0 (t) − αN0 (t)
Ṅ1 (t) = 1−2µ1 u(t) λ1 N1 (t)−(b1 +d1 )N1 (t)
+d1 N1 (t),
+d2 N2 (t) + α N0 (t),
Ṅ1 (t) = 1 − 2u1 (t) λ1 N1 (t) − (b1 + d1 )N1 (t)
.
..
+d2 N2 (t) + αN0 (t),
Ṅl−1 (t) = 1 − 2µl−1 u(t) λl−1 Nl−1 (t)
..
. (5)
−(bl−1 + dl−1 )Nl−1 (t) + dl Nl (t)
(3)
Ṅl−1 (t) = 1 − 2ul−1 (t) λl−1 Nl−1 (t)
+bl−2 Nl−2 (t),
−(bl−1 + dl−1 )Nl−1 (t) + dl Nl (t)
..
.
+bl−2 Nl−2 (t),
Ṅi (t) = λNi (t) − (b + d)Ni (t) + dNi+1 (t)
..
.
+bNi−1 (t) for i ≥ l,
Ṅi (t) = λNi (t) − (b + d)Ni (t) + dNi+1 (t)
+bNi−1 (t) for i ≥ l. where the µi ’s satysfying 0 ≤ µi ≤ 1 are “efficiency fac-
tors”, determining the effectiveness of the drug in relation
to a particular type of cell. Due to the general assumptions
Moreover, multivariable control is allowed, meaning about the model, which were presented at the beginning of
that either some types of the resistant cells can be affected this section, these factors satisfy the following relations:
by chemotherapy or different drugs are being used. A jus-
tification of its usefulness is presented in the following 0 ≤ µi ≤ µi−1 ≤ 1, i = 1, 2, . . . , l − 1. (6)
sections.
Several control problems arising in all these cases
can be addressed based on the model. One of them is 4. Multidrug Protocols
establishing constant control u (in that case it leads to
the determination of feedback parameters) that stabilises Most of the existing forms of therapies consist in using
the infinite dimensional system. In biological terms, it several drugs instead of a single one. Coldman and Goldie
refers to calculating a constant dose of the chemothera- (1986) suggested that a such chemotherapy might reduce
peutic agent that suppresses the growth of the resistant the drug resistance effects. Then the modelling should
subpopulation. However, the constant treatment protocol, take into account the increasing drug resistance to each
which guarantees the decay of the cancer population af- of the chemotherapeutic agents used. There is no known
ter a sufficiently long time, is not realistic. First of all, it mathematical approach to analysing such a problem.
does not take into account the cumulated negative effect of Let us consider the case of a simultaneous use of two
the drug upon normal tissues. To make the solution more types of drugs. Assume the simplest case, in which the
300 J. Śmieja and A. Świerniak
resistance of the cells means that they are insensitive to Although in the mathematical model given above it is
the drugs’ actions, and there are no differences between assumed that the mechanism for resistance is independent
the parameters of cells of different types. Then we could for each drug, the methodology developed makes it pos-
distinguish four different subpopulations of cells: Type 0, sible to analyse also the case when different drugs affect
which is sensitive to both drugs, Type 1 and Type 2, sen- the same gene simultaneously. In that case, the model (7)
sitive only to the first and the second agent, respectively, should be combined with (5). In fact, this combination is
and Type i ≥ 3, resistant to both drugs. The second hy- possible in the general model introduced in Section 6.
pothesis for the original model has to be modified in the
following way:
• A cell of Type 0 may mutate in a short time inter-
val (t, t + dt) into a Type 1 cell with probabilities 5. Phase-Specific Control of the Drug-
α13 dt + o(dt) or into a Type 2 cell with probability Sensitive Cancer Population
α02 dt + o(dt) or into a Type 3 cell with probability
α03 dt + o(dt). The cell cycle is composed of a sequence of phases un-
dergone by each cell from its birth to division. Actu-
• Each cell of Type 1 or Type 2 may mutate in a short
ally, each drug affects the cell being in a particular phase
time interval (t, t + dt) into a Type 3 cell with
and it makes sense to combine these drugs so that their
probabilities α13 dt + o(dt) and α23 dt + o(dt), re-
cumulative effect on the cancer population is the great-
spectively, or into a Type 0 cell with probabilities
est. So far, phase-specific chemotherapy has been con-
d10 dt + o(dt) and d20 dt + o(dt), respectively.
sidered only in the finite-dimensional case, without any
• A cell of Type 3 may mutate in a short time interval regard to problems stemming from increasing drug resis-
(t, t + dt) into a Type 0, Type 1 or a Type 2 cell tance (Świerniak et al., 1997a; Swan, 1990). Combining
with probabilities d30 dt+o(dt), d31 dt+o(dt) and the infinite dimensional model of drug resistance with the
d32 dt+o(dt), respectively, or into a Type 4 cell with phase-specific model of chemotherapy should bring math-
probability b dt + o(dt). ematical modelling much closer to its clinical application.
• A cell of Type i, i ≥ 4, may mutate in a short time Once again, some modification of the assumptions
interval (t, t + dt) into a Type i + 1 cell with prob- underlying the mathematical model presented at the be-
ability b dt + o(dt) and into Type i − 1 cell with ginning of this section should be introduced. The sensitive
probability d dt + o(dt), where αlϕ is several or- subpopulation consists of two types of cells: Type 0, being
ders of magnitude smaller than b and d. in the phase G1 +S, and Type 1, being in the phase G2 M .
The phase-specific drug affects only cells of Type 1. Then
A graph illustrating the possible transfers between the following set of equations can represent the system
different subpopulations is presented in Fig. 1(b). The fol- dynamics:
lowing set of equations is obtained:
Ṅ0 (t) = −λ0 N0 (t) + 2(1 − γ)[1 − u(t)]λ1 N1 (t)
Ṅ0 (t) = 1 − β0 u0 (t) − β1 u1 (t) λN0 (t)
+dN2 (t),
−(α01 + α02 + α02 )N0 (t) + d10 N1 (t)
+d20 N2 (t), Ṅ1 (t) = −λ1 N1 (t) + λ 0 N0 (t),
Ṅ1 (t) = 1 − 2u0 (t) λN1 (t) − (α13 + d10 )N1 (t) Ṅ2 (t) = λ2 N2 (t) − (b + d)N1 (t)
(8)
+d31 N3 (t) + α01 N0 (t), +2γ 1 − u(t) λ1 N1 (t) + bN3 (t),
..
Ṅ2 (t) = 1 − 2u1 (t) λN2 (t)
.
−(α23 + d20 )N2 (t) + d32 N3 (t) (7) λNi (t) − (b + d)Ni (t) + dNi+1 (t)
Ṅ i (t) =
+α02 N0 (t), +bNi−1 (t) for i ≥ 3,
Ṅ3 (t) = λN3 (t) − (b + d30 + d31 + d32 )N3 (t)
where γ is the probability of the primary mutational
+dN4 (t) + α13 N1 (t) + α23 N2 (t),
event. A graph illustrating possible transfers between dif-
.
..
ferent subpopulations is presented in Fig. 1(c). Similarly,
Ṅi (t) = λNi (t) − (b + d)Ni (t) + dNi+1 (t) a multidrug therapy including blocking drugs (Świerniak
et al., 1997b; Brown and Thompson, 1975) as well as the
+bNi−1 (t) for i ≥ 4,
killing agent could be analysed in the same way, as pre-
where β1 and β2 are efficiency factors, and β0 +β1 ≤ 2. sented in the subsequent sections.
Different models of chemotherapy taking into account drug resistance stemming from gene amplification 301
0 2 .. Ã2 l −1 m l −1
N l (t )
N l −2 = ∑∑b j
0 ,i u i (t ) N j (t ) + ∑a l − 2 , i N i (t ) N l −1 (t )
j =0 i =0 i =0
l −1 m l
. ∑∑b j
∑a
N l −1 = 0 ,i u i (t ) N j (t ) + l −1, i N i (t )
B̃ i .. 0 1 j =0 i =0 i =0
B = . . . . . .· . . . . . , (11)
03
a00 a01 ··· a0,l−1 0
a10 a11 ··· a1,l−1 0
N l (t ) = c1 N l −1 (t ) + a2 N l (t ) + a3 Nl +1 (t )
Ã1 = .. .. .. , N l +1 (t ) = a1 N l (t ) + a2 N l +1 (t ) + a3 N l + 2 (t )
···
0
. . .
c1 a2 a3 0 0 0 ··· Fig. 2. Decomposition of the system model.
0 a1 a2 a3 0 0 ···
Ã2 = ,
0 0 a1 a2 a3 0 ···
.. .. .. .. .. .. .. 7.1. Infinite Dimensional Subsystem
. . . . . . .
First, consider the infinite dimensional tail without the in-
fluence of cells Nl−1 :
bi0,0 bi0,1 ... bi0,l−1
i
b1,0 bi1,1 ... bi1,l−1
Ṅl (t) = a2 Nl (t) + a3 Nl+1 (t),
B̃ i = ,
.. .. ..
Ṅ (t) = a1 Nl (t) + a2 Nl+1 (t) + a3 Nl+2 (t),
l+1
...
. . .
i
..
bl−1,0 bil−1,1 ... bil−1,l−1 . (12)
Ṅi (t) = a1 Ni−1 (t) + a2 Ni (t) + a3 Ni+1 (t),
..
u(t) stands for the m-dimensional control vector u =
.
[u0 u1 u2 . . . um−1 ]T , 01 , 02 and 03 are zero matri-
ces of dimensions l × ∞, ∞ × (l + 1) and ∞ × ∞, Using methods similar to that outlined in our previ-
respectively, l > m. ous works devoted to biomedical modelling (Świerniak
It is important to note that the model parameters sat- et al., 1999; Śmieja et al., 2000), it is possible to show
isfy the relations a3 > a1 > 0 and a2 < 0. However, a that for the initial condition Ni (0) = δik (the Kronecker
full problem analysis can be made in other possible cases delta), i.e., Nk (0) = 1, Ni (0) = 0 for i 6= k, the follow-
(e.g., when no additional conditions are to be satisfied by ing relations hold true:
the parameters a1 and a3 ) using exactly the same line p !k−l+1
of reasoning. The performance index to be minimised is k 1 s−a2 − (s−a2 )2 −4a1 a3
Nl (s) = , (13)
given by (4). a3 2a1
302 J. Śmieja and A. Świerniak
(the superscript k is introduced to emphasize the index of Then, applying the standard control theory techniques
the state variable with a non-zero initial condition). Now, (Zadeh and Desoer, 1963), the following relation holds
assume that k = l. Then, after calculating the inverse true for u(t) = 0:
Laplace transform, the following formulae are obtained:
√ Xl−1 (s)
1
r
a3 I1 2 a1 a3 t
K2 (s) = = C(sI − Ã1 )−1 B̂ 1 . (22)
l
Nl (t) = exp(a2 t), (15) Xl (s)
a3 a1 t
X Taking into account the linear form of this system, it
NΣl (t) =
Ni (t) = exp (a1 + a2 + a3 )t is possible to present the model in the form of a block di-
i̇≥l agram, shown in Fig. 3. This makes it possible to analyse
r Z t √ the dynamical properties of the closed-loop system.
a3 I1 (2 a1 a2 τ )
× 1−
a1 τ
0
Nl (s) Nl−1(s)
K2(s)
× exp − (a1 + a3 )τ dτ , (16)
will be considered in the next section. Instead, a model The goal is to minimise the performance index given
transformation into an integro-differential one is proposed by (4). Due to the particular form of both the performance
in this section. index and the equation governing the model, it is possible
Set to find the solution to the problem applying the appropri-
N0 ate version of Pontryagin’s maximum principle (Pontrya-
. gin et al., 1962).
x̃ = . (24)
. It is important to notice that, although the perfor-
Nl−1 mance index (4) seems to consist of two components
and C k = [cj ], ck = 1, cj = 0 for j 6= k, i = (a sum and an integral), the sum actually involves another
1, 2, . . . , l − 1. Let us also assume that Ni (0) = 0 for integral, which stems from (19) and (20). Therefore, it
i > l − 1 (once again it should be stressed that any finite should be rewritten to emphasise this relation:
non-zero initial condition can be incorporated into the fi- l−1
X
nal solution). Then the last equation in the first subsystem, J= Ni (T ) + r1 NΣl (T )
influenced directly by control, as presented in Fig. 2, can i=0
Z T m
be transformed into the integro-differential form l
X
+ r1 c1 NΣ (T −τ )Nl−1 (τ )+r uk (τ ) dτ . (30)
l−1 X
m l−1 0 k=0
X X
Ṅl−1 (t) = bjl−1,i ui (t)Nj (t) + al−1,i Ni (t) A number of formulations of necessary conditions
j=0 i=0 i=0
for the optimisation problem for dynamical systems gov-
Z t erned by integro-differential equations can be found in the
+ al−1,l k1 (t − τ )Nl−1 (τ ) dτ , (25) literature, e.g., (Bate, 1969; Connor, 1972; Gabasov and
0 Kirilowa, 1971). However, they are usually too general to
where k1 (t) is the inverse Laplace transform of K1 (s) be efficiently applied to such a particular problem or they
given by (18). have too strong constraints (e.g., regarding the smooth-
ness of the control function). Nevertheless, following the
Similarly, other equations can also be rewritten in the
line of reasoning presented by Bate (1969), it is possible
same way, leading to the transformation of the model (9)
to derive the necessary conditions for optimal control:
into the following form:
X m
opt
uk (t) + pT (t)h(u, x̃)
Z t
u (t) = arg min r
˙˜x = h(u, x̃) + f˜(x̃, t, τ ) dτ , x̃(0) = x̃0 , (26) u
k=0
0 Z T
where h(·, ·) are f˜(·, ·, ·) are the respective l-dimension- +al−1,l pl−1 (τ )k1 (t−τ )Nl−1 (τ ) dτ , (31)
t
al vector functions:
l−1 X
X m l−1
X ṗT (t) = − q T (t) + pT (t)hx̃ (u, x̃)
hk (u, x̃) = bjk,i ui (t)Nj (t) + ak,i Ni , (27)
j=0 i=0 i=0
Z T
+ T ˜
p (τ )fx̃ (t − τ )dτ , (32)
( t
0 for k < l−1,
f˜k (x̃, t, τ ) = (28) T
al−1,l k1 (t−τ )Nl−1 (t) for k = l−1. q(t) = 0 . . . 0 r1 c1 NΣl (T − t) ,
pi (T ) = 1, i = 0, 1, . . . , l − 1, (33)
9. Necessary Conditions for Optimal Control p(t) being the adjoint vector.
Taking into account the constraint (29) and the bi-
After the transformation of the system description pre- linear form of (27), it can be proved that, in order to sat-
sented in the previous section, it is possible to effectively isfy (31), optimal control must be of the bang-bang type.
address the resulting optimal control problem. Then, to find an optimal number of switches and switch-
Let the system be governed by Eqn. (9), which is then ing times, a gradient method can be developed, following
transformed into the form (26). The control is bounded, the line of reasoning presented in (Śmieja et al., 1999).
i.e.,
0 ≤ uk (t) ≤ 1, (29)
10. Conclusions
where uk (t) = 1 represents the maximum allowable dose
of the drug k and uk (t) = 0 represents no application of This paper is concerned with an infinite dimensional bi-
the drug k. linear model of dynamical systems. Based on model de-
304 J. Śmieja and A. Świerniak
composition, it is possible to analyse analytically some of Connor M.A. (1972): Optimal control of systems represented by
their dynamical properties. The transformation of system differential-integral equations. — IEEE Trans. Automat.
description into one integro-differential equation allows Contr., Vol. AC-17, No. 1, pp. 164–166.
solving an optimal control problem with the performance Curtain R.F. and Zwart H.J. (1995): An Introduction to Infinite-
index defined in the l1 space of summable sequences. Dimensional Linear Systems Theory. — New York:
One of the possible model applications is the mod- Springer.
elling of drug resistance emergence in cancer cells and Gabasov R. and Kirilowa F.M. (1971): Qualitative theory of op-
the analysis of possible chemotherapy protocols. Until timal processes. — Moscow: Nauka, (in Russian).
now, treatment protocols have been designed mainly on
the basis of experimental results and the general knowl- Harnevo L.E. and Agur Z. (1993): Use of mathematical models
edge about the drug activity. However, no general math- for understanding the dynamics of gene amplification. —
ematical approach exists which would help to explain the Mutat. Res., Vol. 292, No. 1, pp. 17–24.
obtained results or to design a treatment in chemotherapy. Kimmel M. and Axelrod D.E. (1990): Mathematical models of
The results of this work can be used to reveal the desired gene amplification with applications to cellular drug re-
form of an optimal treatment protocol and to give some sistance and tumorigenicity. — Genetics, Vol. 125, No. 3,
hints as to its development. It can be also employed in pp. 633–644.
a qualitative analysis of the chosen protocol, taking into Kleinrock L. (1976): Queuing Systems. Vol. 1: Theory. — New
account both the increasing resistance to chemoterapeu- York: Wiley.
tic agents and allowing multidrug and phase-specific treat-
ments. Olofsson P. and Kimmel M. (1999): Stochastic models of telom-
ere shortening. — Math. Biosci., Vol. 158, No. 1, pp. 75–
The main application of the obtained results is in the 92.
field of biomedical modelling, from which the form of the
performance index in the analysed problem stems. Never- Polański A., Kimmel M. and Świerniak A. (1997): Qualitative
theless, taking into account the broad class of systems de- analysis of the infinite-dimensional model of evolution of
scribed by the presented model, it should easily be adapted drug resistance, In: Advances in Mathematical Population
Dynamics—Molecules, Cells and Man (O. Arino, D. Ax-
for usage in other areas.
elrod and M. Kimmel, Eds.). — Singapore: World Scien-
tific, pp. 595–612.
Acknowledgement Pontryagin L.S., Boltyanski V.G., Gamkrelidze R.V. and Mis-
chenko E.F. (1962): Mathematical Theory of Optimal Pro-
This work was partially supported by the grant of the cesses. — New York: Wiley.
State Committee for Scientific Research (KBN) in Poland
Ramel C. (1997): Mini- and microsatellites. — Environmental
No. 4T11F 01824.
Health Perspectives, Vol. 105, Suppl. 104, pp. 781–789.
Śmieja J., Duda Z. and Świerniak A. (1999): Optimal control
References for the model of drug resistance resulting from gene am-
plification. — Prep. 14th IFAC World Congress, Beijing,
Arino O., Kimmel M. and Webb G.F. (1995): Mathematical China, Vol. L, pp. 71–75.
modelling of the loss of telomere sequences. — J. Theor.
Biol., Vol. 177, No. 1, pp. 45–57. Śmieja J., Świerniak A. and Duda Z. (2000): Gradient method
for finding optimal scheduling in infinite dimensional mod-
Axelrod D.E., Baggerly K.A. and Kimmel M. (1994): Gene am- els of chemotherapy. — J. Theor. Med., Vol. 3, No. 1,
plification by unequal chromatid exchange: Probabilis- pp. 25–36.
tic modelling and analysis of drug resistance data. — J.
Theor. Biol., Vol. 168, No. 2, pp. 151–159. Swan G.W. (1990): Role of optimal control theory in cancer
chemotherapy. — Math. Biosci., Vol. 101, No. 2, pp. 237–
Bate R.R. (1969): The optimal control of systems with transport 284.
lag, In: Advances in Control and Dynamic Systems (C.T.
Leondes, Ed.). — Academic Press, Vol. 7, pp. 165–224. Świerniak A., Śmieja J., Rzeszowska-Wolny J. and Kimmel
M. (2001): Random branching walk models arising in
Brown B.W. and Thompson J.R. (1975): A rationale for syn- molecular biology—control theoretic approach. — Proc.
chrony strategies in chemotherapy, In: Epidemiology IASTED MIC Conf., Innsbruck, Austria, Vol. II, pp. 584–
(D. Ludwig and K.L. Cooke, Eds.). — Philadelphia: 589.
SIAM Publ., pp. 31–48.
Świerniak A., Kimmel M., Polański A. and Śmieja J. (1997a):
Coldman A.J. and Goldie J.H. (1986): A stochastic model for the Asymptotic properties of infinite dimensional model of
origin and treatment of tumors containing drug-resistant drug resistance evolution. — Proc. ECC’97, Brussels, TU-
cells. — Bull. Math. Biol., Vol. 48, No. 3–4, pp. 279–292. A-C-4, CD-ROM.
Different models of chemotherapy taking into account drug resistance stemming from gene amplification 305
Świerniak A., Polański A., Duda Z. and Kimmel M. (1997b): Świerniak A., Polański A., Śmieja J., Kimmel M. and
Phase-specific chemotherapy of cancer: Optimisation of Rzeszowska-Wolny J. (2002): Control theoretic approach
scheduling and rationale for periodic protocols. — Biocy- to random branching walk models arising in molecular bi-
bern. Biomed. Eng., Vol. 16, No. 1–2, pp. 13–43. ology. — Proc. ACC Conf., Anchorage, pp. 3449–3453.
Świerniak A., Kimmel M. and Polański A. (1998): Infinite di- Zadeh L.A. and Desoer C.A. (1963): Linear System Theory. The
mensional model of evolution of drug resistance of can- State Space Approach. — New York: McGraw-Hill.
cer cells. — J. Math. Syst. Estim. Contr., Vol. 8, No. 1,
pp. 1–17.